G. F. Malykhina, A. V. Merkusheva
Saint-Petersburg
The limitedness of the signal local-stationarity
concept in information measurement and information control systems (ICS)
reflecting the state of the controlled plant (subsystem) demands the use
of more perfect analysis and processing methods, time-frequency transformations
and algorithms for neural networks (NN). Significant problems occur when
the plant (subsystem) state control is implemented in conditions where
some parameters have no effect on the measuring system sensors, i.e. in
conditions of incomplete information. The solution of this problem is obtained
based on the analysis of plant-ICS system dynamics equations (in the state
parameter space) and on the use of temporal NN algorithms. The first
(of three) paper parts discusses the NN structure and learning algorithms
that may adequately represent the data and controlled process dynamics.
NN structures are analyzed on the basis of a neural filter concept and
learning -- on the basis of a time-dependent back-propagation algorithm.